Hacker News Reader

Comment by onlyrealcuzzo | Hacker News Reader

As someone who works in Data Science at FAANG - if you look hard enough - there is something questionably wrong in every step of the data funnel.

And that's when I believe people do have a somewhat best effort to maximize profits. There are plenty of people that only care about career progression and think they can get away with lying and cheating their way to the top. They wouldn't believe that if it didn't work sometimes.

These medical studies are also run mainly to maximize profits, also by some career climbers. They are not run virtuously for the betterment of society.

So I would be astounded if they are as reliable as people might like to believe.

Maybe I'm just being grossly skeptical. Actually, I'd feel better if someone could convince me I'm completely unfounded here.

View on HN

"And that's when I believe people do have a somewhat best effort to maximize profits."

Nope. I actually think that if you do scientific research as a company (profit) it may make you less bad/less likely to do fraud compared to academia (non-profit).

Reason is that there are more ways to punish you, employees, board, investors, etc in a profit seeking vehicle, and as a profit seeking vehicle being caught must be part of the profit seeking calculation – in the end, the world of reality/physics will weigh your contribution.

I believe there is evidence that there is more fraudulent scientific research happening in non-profit vehicles/academia. Take for example an area where there are fewer profit seeking companies participating - social sciences. It's dominated by academia. Now look at the replication rate of social sciences.

All three points in last paragraph seem wholy unrelated to each other and your larger point. I agreed with first half.

I have a simpler reason for the same belief.

You can fake everything except a well designed A/B test. At FAANG scale, a statistically significant A/B test requirement will stop the worst fraud before it hits the user.

And also - you can somewhat take care of bugs by evenly distributing them to your test & control group.

FAANG engineers: “hold my beer”

Seriously though, as a person who has built related systems at FAANG, yes this problem exists there. Your beautiful cathedral of an A/B testing framework is covered in knobs that are just perfect for p-hacking.

I used to work data entry as a research assistant, for some non-profit academia.

We saw flaws in the data collection - basically the people tasked to collect data were being lazy and some were making stuff up. We know the made up stuff when we see it. Outliers are fine and some groups do better than expected, but an entire group from one data collecter shouldn't be 100% outliers.

But we had to enter the data anyway. We were told to smooth the bad data to what was expected. So the outliers that were low were smoothed high, the high outliers were left alone because they seemed right. But those of us who were spending hundreds of hours on data entry had an intuitive feel of what an outlier looked like.

IMO everything should have been entered as is and the computer data would just be filtered out if it was deemed from a corrupt source. But the data in the computer was biased to match what the research wanted to prove.

So I agree that non-profits can be corrupt too, just because of the incentives each part of the way. We were being paid about half a cent per column of data. So some assistants were lazy and filling in data that could be right, or skimming on fields like address which are longer and less likely to be flagged.

Exactly this.

So many people have the opinion that private research must be flawed because of the profit motive, but the profit motive ensures that someone will be motivated to take oversight seriously, and have the power to punish misbehavior.

Free markets, as ugly as they are sometimes, are still the best way we have of ensuring that incentives align with outcomes.

You might like https://gwern.net/backstop

War also works, to a certain extent, as a source of truth to align incentives with outcomes. But it's horribly expensive even in terms of economics alone, not to mention the human tragedy.

Luckily free markets work as a backstop, too. Add in free movement of people (who often want to come to better run places), and free movement of capital, and you have a winning combination.

Another stroke of luck: even if you only implement a very partial version of 'free', you still get partial benefits. Slightly freer markets are typically slightly more efficient. It's not an all or nothing proposition.

But work in social sciences will often be inherently harder to replicate due to, for example, groups of human beings being more complex than groups of proteins - so there are a lot more variables, confounding factors, difficulties doing double-blind randomised trials, and so on...

Yes, but that's why the companies like Google or Facebook spend so much time and effort on their social science experiments.

Granted, they are mostly interested in a very small sliver of social science: 'how can you get people to directly or indirectly spend more time online and look at more ads'; but they are very, very interested in getting robust results that replicate well. They are also interested in figuring out how the results vary between different cultures and over time.

I think it's true that industry is more likely to produce reliable research compared to academia, but for a different reason.

In academia, you essentially have the student who does the work, the professor, and the person who funded the grant, and that's essentially the sum total of people supervising the data. There's not a lot of people to call you out on fudged (or outright faked) data, and all of them are likely to be very invested in the success of the research.

Turn to industry, and you have a similar set of people--the worker, the manager, and the head of the research department--except maybe a few more levels of manager (depending on the scale of the project). But since the goal is usually productization in industrial research, you usually have to turn to the product divisions and convince their executive chain as well of the merits of your research. And unlike everybody else mentioned so far, this group of people isn't invested in the success of the research. You might even be competing against other research teams that have different alternatives, and those people are going to be actively invested in the failure of your research so that their research makes it instead.

Also the guys in the product team are interested in whether they can actually reproduce eg your novel synthesis method for your favourite molecule on a large scale.

The 'product' of academic research is a published paper. The product of industrial research is an actual product.

(This does not apply when your research is about eg effectiveness of a new drug. The product people can sell that drug on the strength of that research. Whether that research replicates or not is only of indirect concern in that case.)

Completely anecdotally, I find this to be true. I do pre-clinical research (disclaimer: for a pharma company), and if you hand me a paper describing some research result in my field, it’s much more likely to reproduce in my hands if it’s from industry. I know from talking to my academic research friends that there are a ton of shortcuts, rotating inexperienced staff, and misaligned incentives. I’m not saying pharma is perfect, but in my experience, they’re more reproducible.

Fwiw, I do not know of any data in my realm which have been molded, cherry picked, intentionally misrepresented, falsified, or otherwise fake or flawed. I don’t work with clinical trial data, so if that were happening, it wouldn’t be on my desk.

I used to work in agricultural R&D in the private sector, and we collaborated with several universities.

The biggest difference I saw was that universities were very short-term focused, while we were more long-term focused.

The universities had a constant churn of personnel, as new PhD candidates appeared and old ones left, whereas our own researchers and technicians stuck around for far longer.

Additionally, they were so hyperfocused on grants and papers that they tended to not put as much effort into replication, since that didn't pay their bills. By comparison, we typically repeated our experiments ad nauseum; it was common for us to perform the same experiment twice a year (once in the northern hemisphere and then again in the southern hemisphere) for a decade or more, gradually iterating and refining our processes along the way.

Even if we initially got negative results, we'd beat that dead horse for a few years to make sure. Occasionally it turned out not to be so dead after all.

>Actually, I'd feel better if someone could convince me I'm completely unfounded here.

Unfortunately, your skepticism is not unfounded. Those in the industry conclude the same. Take, for instance, the editor in chief of The Lancet:

>The case against science is straightforward: much of the scientific literature, perhaps half, may simply be untrue. Afflicted by studies with small sample sizes, tiny effects, invalid exploratory analyses, and flagrant conflicts of interest, together with an obsession for pursuing fashionable trends of dubious importance, science has taken a turn towards darkness.

https://www.thelancet.com/journals/lancet/article/PIIS0140-6...

I’ve literally never seen data-driven business decisions that weren’t using fatally flawed datasets or methods so bad that you’d be a fool to believe you were getting anything but gibberish out of them, except in trivial cases.

You quickly learn not to be the guy pointing out the problem that means we’ll need several people and months or years to gather and analyze data that would allow them to (maybe) support or disprove their conclusion, though. Nobody wants to hear it… because they don’t actually care, they just want to present themselves as doing data-driven decision making, for reasons of ego or for (personal, or company) marketing. It’s all gut feelings and big personalities pushing companies this way and that, once you cut through the pretend-science shit.

“Yeah, that graph looks great (soul dies a little) let’s do it”

>Nobody wants to hear it… because they don’t actually care, they just want to present themselves as doing data-driven decision making

Oh, that hits home

I know someone who’s fabricated almost all their work. Their lab leads always cover it up and don’t retract their work. The crazy part is this person has failed upwards and is currently a senior researcher at a major research institution. They’ve been at UCLA and UCSF for medical research.

Absolutely insane to watch.

Objectivity and honesty can be hard to find if all someone cares about is their reputation as a competent researcher or climbing the ladder. What do you think a potential solution would be for this? I feel like even in my own experience, trying something out and it not working feels like failure, when in fact to proclaim it a success or "fix" it is truly what harms both the endeavor for truth and the people reliant on the outcomes of these surveys

> Objectivity and honesty can be hard to find if all someone cares about is their reputation as a competent researcher or climbing the ladder.

It seems like destroying the reputation and career of people who fake science would be a great start. If you're willing to fake data and lie to get results, there will always be an industry who'd love to hire you no matter how tarnished your reputation is. We need a better means to hold researchers accountable and we need to stop putting any amount of faith in any research that hasn't been independently verified through replication.

Today the lobby for orange juice manufactures can pay a scientist to fake research which shows that drinking orange juice makes you more attractive, and then pay publications to broadcast that headline to the world to increase sales. We should have some means to hold publications responsible for this as well.

> destroying the reputation and career of people who fake science would be a great start

When so many reports are faulty and fraudulent, that might instead be the great start of destroying the careers of those who would have revealed the fraudulent research?

I wonder what'd happen if researchers got compensated and funding based on other things, unrelated to papers published. But what would that be

> I wonder what'd happen if researchers got compensated and funding based on other things, unrelated to papers published. But what would that be

See what they do in the parts of industry where they need their research to work.

Eg how do battery manufacturers compensate and incentives their researchers that are aiming to improve various characteristics of batteries? How do steel mills manage and reward their metallurgists? How does Intel's research work?

Aha, so they get paid the same, regardless of if their research works or not. And possibly some bonuses if things go well for the company? And if they all the time do nothing useful and manipulate data, they get fired?

But that's different from a PhD student -- they're not embedded in any organization that would notice if the research works or not?

Maybe if the universities partnered somehow with different companies, and the researchers got extra compensation if a company decided to make real world use of the research?

(On top of some base salary)

But who would determine if a company had made use of a certain research paper? What would the company gain, by keeping track and reporting back? Maybe more good research

(but I'd guess few companies would be that much forward-looking?)

> Aha, so they get paid the same, regardless of if their research works or not.

I don't know. Is that speculation on your part, or something you figured out?

> But who would determine if a company had made use of a certain research paper? What would the company gain, by keeping track and reporting back? Maybe more good research

> (but I'd guess few companies would be that much forward-looking?)

That's why I am saying we should look what real companies are actually doing already in reality. We might have to leave our armchairs for that.

> Is that speculation on your part, or something you figured out

That's just normal monthly wages, how things usually work.

> look what real companies are actually doing already

But you can't look at what companies are doing now, to find out if new research is useful? The companies can't yet have started doing the things that any new & good research enables (since it wasn't known before).

Could take years until they make use of the research

(But maybe you meant something else)

I mean look at what companies have been doing in the past to organise and incentivise research, and see what worked.

Look at what processes worked (and didn't work!). Not at what specific inventions worked.

If you have any books or articles you like related to that, it'd be interesting to hear :-)

(Research by Google about psychological safety comes to my mind.)

I don't think that's skeptical.

I do think there is an even worse issue - which is funding. The money incentive means you can fund studies that support whatever you want.

Really curious how a profitable industry like medical research ranks against less profitible.

Reply on news.ycombinator.com